You answer "what's the delivery time?" twenty times a week, and you usually do it at night, after the real workday is over. An AI customer service agent is a program that reads a customer's question, looks up the answer in your own information (FAQ, terms of sale, catalog, order tracking) and replies on its own, 24 hours a day. For a small business handling 200 to 500 messages a month, the technical bill usually lands between €5 and €40 per month. The real cost is elsewhere: the setup time, the upkeep of your knowledge base, and the requests the machine absolutely has to hand off to a human. Below you'll find the numbers line by line, three ways to set it up, and an honest list of what breaks in production.
What an AI customer service agent does, and what it doesn't
An AI agent handles questions whose answer is already written down somewhere, and handles anything requiring a business decision badly.
What it does well from week one:
- delivery times, shipping costs, areas served
- opening hours, address, how to book an appointment
- the status of an order, provided it's connected to your sales tool
- product questions: material, size, compatibility, warranty, care
- triaging requests: it qualifies the message and files it in the right queue before you even open it
What it doesn't handle, and where you need a human fallback:
- a refund outside your stated terms, a goodwill gesture, a price negotiation
- an angry customer threatening to leave a one-star review
- anything touching a person's health, legal situation or safety
- a request whose answer exists nowhere in your documents: it then has to say it doesn't know and hand off
The difference from a classic button-based chatbot is simple: the agent understands a question phrased any which way, typos and vague vocabulary included, and it can trigger actions (look up an order, create a ticket, send a confirmation email). If you want to see what building one looks like step by step, the guide on creating a chatbot for your business with Claude covers the technical side.
How much an AI customer service agent costs a small business
Budget €2 to €7 per month in model calls for 300 conversations, €0 to €55 in surrounding tools, and 6 to 15 hours of work the first time around.
AI models are billed per token. A token is roughly three quarters of a word. You pay for the tokens you send in (the customer's question, your FAQ, the conversation history) and the ones the model produces. According to the public Claude API pricing, a lightweight model like Haiku sits around $1 per million input tokens and $5 for output; a top-tier model around $3 and $15.
Here's the concrete math for an average six-message conversation, with your FAQ injected on every call: roughly 4,000 cumulative input tokens and 600 output tokens. With the lightweight model that comes to $0.007, less than a cent. With the top-tier model, about 2 cents. Across 300 conversations a month, you're looking at €2 to €7.
| Line item | Ballpark for 300 conversations/month |
|---|---|
| AI model calls | €2 to €7 |
| Hosting the automation (n8n cloud or a small server) | €0 to €25 |
| Chat widget on your site | €0 to €30 |
| Upskilling | Monthly subscription with Skilzy |
| Your setup time | 6 to 15 hrs, one time only |
| Monthly upkeep | 1 to 2 hrs |
The line item that really blows up a budget is the per-resolution pricing some support platforms use: around €1 per conversation resolved by the AI. On 300 conversations with half resolved automatically, you land at €150 a month, versus €5 if you call the model directly. It's not a rip-off: you're paying for the interface, the hosting, the analytics and the fact that you don't have to build anything. But know what you're buying.
One last number to keep in mind: your own time. If you spend 20 hours a month answering repetitive messages and you value your hour at €40, your current customer service costs €800 a month. That's the figure to compare against, not zero.
Three ways to set up your AI agent
The most profitable approach over twelve months is the one where you learn to build the agent yourself; the fastest is the one where you rent a ready-made tool.
Method 1: learn to build it and keep it in-house
You assemble three building blocks: an AI model called through an API, a knowledge base (your product pages, your FAQ, your terms of sale) and an automation tool like n8n that wires it all up to your inbox, your site or your WhatsApp. No code to write: n8n works with blocks you connect together, and the official n8n documentation covers every block.
The sticking point for a beginner is never technical, it's the method: what to put in the system prompt, how to force the agent to say "I'm handing this to a human" instead of making things up, how to test on 50 real questions before going live. The Skilzy freelance AI program covers exactly this chain, from writing reliable prompts to shipping an automation that holds up in production. It's designed for people who want to bill this work to clients, and it's the same skill you need for your own business.
Two practical details. First, the included AI Lab lets you test the tools with credits provided, instead of stacking up three €20 subscriptions while you learn. Second, two state-recognized certifications (RS7439 in AI content marketing, RS6792 in AI and sales) are eligible for funding if you go through a training scheme.
Realistic budget: €30 a month for training over two to three months, then €5 to €30 a month to run. You stay in control of the agent and can tweak it on a Sunday night without waiting on a contractor.
Method 2: an off-the-shelf support tool with built-in AI
Chat and ticketing software all now offer an AI layer that reads your help center articles and answers in your place. You drop a script on your site, import your FAQ, and you're live in two hours.
Upsides: zero building, clean interface, analytics included, human escalation handled natively, compliance documented by the vendor.
Downsides: pricing runs from around twenty euros per month per seat up to roughly €1 per AI resolution, which climbs fast with volume. You're also limited to what the vendor planned for: connecting your homemade inventory spreadsheet or your regional POS software is often impossible. Check current price sheets before committing, as they change several times a year.
Who it's for: you have a standard e-commerce site, steady volume, and zero desire to touch anything technical.
Method 3: have a custom agent built
You hand the project to a freelancer or an agency. Quotes in this space run from €1,500 for a FAQ agent wired to a brochure site up to €8,000 and beyond for an agent connected to an ERP, with customer authentication and real-time order tracking.
Upsides: it works, and fast, without you learning a thing.
Downsides: every change goes through the provider, and a customer service agent changes constantly (new product, new delivery time, new promo). Negotiate access to the automation editor into the quote from the start, plus a training session so you can update the knowledge base yourself. Without that, you'll pay €150 in billable work to change one sentence about shipping costs.
| Method | Year 1 cost | Time to launch | Autonomy |
|---|---|---|---|
| 1. Build it after training | €150 to €500 | 2 to 6 weeks | Total |
| 2. Off-the-shelf tool | €300 to €2,000 | 2 hours | Low |
| 3. Custom development | €1,500 to €8,000 | 3 to 8 weeks | Medium |
The real limits, the ones you discover after going live
An AI customer service agent fails mainly on three fronts: it invents things when it doesn't know, its knowledge base goes stale, and it's subject to legal obligations nobody reads.
It makes things up. A language model produces the most probable continuation of a text, not the truth. Ask it whether your product is compatible with a reference it has never seen, and it will sometimes answer "yes, absolutely" with total confidence. The defense is in the system prompt: an explicit ban on answering outside the documents provided, and a mandatory fallback sentence. Test that instruction with twenty trick questions before opening it to the public.
Its knowledge base rots. You switch carriers, you raise shipping to €4.90, you discontinue a product. If nobody updates the documents, the agent will quote a price that no longer exists for six months, and you'll find out through a customer review. Block one hour a month in your calendar for that review pass — it's the main hidden cost.
It's legally regulated. Since August 2026, the transparency obligations of the EU AI Act have applied: the person has to know they're talking to a machine. A clear welcome line is enough. On the personal data side, your conversations contain names, addresses, sometimes order numbers: you need a legal basis, a retention period, a mention in your privacy policy, and a check on where your provider hosts the data. The CNIL's recommendations on AI lay out the framework that applies in France.
It can't break out of a loop. The worst scenario for your reputation isn't a bad answer, it's the stuck customer who can't find any way to reach a human. Put a visible handoff button on every message, and trigger escalation automatically after two unsatisfying replies or as soon as a word like "refund", "lawyer" or "cancel" shows up.
Its tone goes off the rails. By default, an AI agent writes like a translated American legal department. Give it three examples of your own past replies in the prompt and the result changes completely. The same principle applies to your written replies in general, as covered in the guide on automating your emails with AI.
Measuring the return and starting out cleanly
Your agent pays off when it closes at least 40% of your conversations with no human involvement, while complaints stay flat.
Three metrics are enough, checked every month:
- Autonomous resolution rate: the share of conversations finished without you. Below 30%, your agent doesn't have enough information. Above 70%, check that it isn't blocking customers who wanted a human.
- Time saved: number of conversations resolved times your average response time. 135 conversations resolved at 4 minutes each works out to 9 hours a month.
- Complaints and reviews: if the number of negative reviews mentioning support goes up, your resolution rate is a false friend.
Your launch plan, in this order:
- Step 1. Export your last 100 customer messages and sort them. You'll see that 6 or 7 questions cover 80% of the volume.
- Step 2. Write the ideal answer to those 7 questions, in your own voice, with exact figures. That's your knowledge base, and it's 80% of the work.
- Step 3. Set the agent up on a single channel, the one with the highest volume (site form, or email).
- Step 4. Keep it in draft mode for two weeks: it prepares the reply, you approve before it goes out. That way you fix mistakes without showing them to a customer.
- Step 5. Switch to fully automatic on simple questions only, and keep human approval on everything else.
This draft-then-approve logic before full automation applies to almost every process in a small business; the guide to automating your business with AI shows how to apply it to invoicing, prospecting and reporting.
Key takeaways
The technical cost of an AI customer service agent has become marginal for a small business: a few euros a month if you build it, a few dozen if you rent it. What you really pay is about ten hours of prep and one hour of upkeep a month. In exchange, you often get 8 to 10 hours back. Start small, on one channel, with seven questions and human approval, and expand only when the numbers look good.
